A sleep quality analysis system based on deep learning

Abstract

The application provides a sleep quality analysis system based on deep learning, and relates to the technical fields of biomedical engineering and artificial intelligence, and comprises a multi-modal signal acquisition module for acquiring physiological signals and environmental signals of a target user; the physiological signals include brain wave signals, electromyography signals and skin resistance signals; the environmental signals include temperature and humidity signals and air quality signals; a denoising module for denoising the physiological signals and the environmental signals to obtain denoised signals; a feature extraction module for extracting features from the denoised signals to obtain signal features; a feature processing module for performing data cleaning, data filling and normalization processing on the signal features to obtain normalized features; an evaluation module for inputting the normalized features into a trained sleep quality evaluation model to obtain an evaluation result; and the sleep quality evaluation model is constructed by a deep learning network. The application can more accurately evaluate sleep quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical engineering and artificial intelligence, in particular to a sleep quality analysis system based on deep learning. BACKGROUND

[0002] With the acceleration of life pace, sleep quality problems are increasingly concerned. Traditional sleep monitoring methods usually rely on polysomnography (PSG), but its device is complex and not convenient for daily use. In recent years, sleep analysis methods based on data collection of wearable devices combined with deep learning technology have gradually become a research hotspot.

[0003] The invention patent with application number 202311397993.6 discloses a user sleep quality management system based on detection data analysis, which comprises: a sleep state monitoring module, including an environment monitoring unit, a physiological monitoring unit and a sleep posture monitoring unit, for monitoring the sleep environment state, physiological parameters and sleep posture state of the user respectively, obtaining monitoring data; an analysis unit for evaluating the sleep state of the user according to the monitoring data and the historical sleep data of the user, obtaining evaluation results; a traceability module for tracing the factors affecting the sleep of the user according to the evaluation results and issuing corresponding prompt suggestions; the sleep posture monitoring unit includes a sensing pillow and a sensing mattress, the sensing pillow includes a pillow body and a plurality of groups of pressure sensors; the sensing mattress includes a mattress body and a plurality of groups of pressure sensors; the process of evaluating the sleep state of the user includes: S1, monitoring the sleep posture change state of the user according to the parameters of the pressure sensors in the sensing pillow and the sensing mattress, obtaining sleep posture state monitoring data; S2, obtaining the physiological parameter values of the user during sleep through the physiological monitoring unit; S3, evaluating the sleep state of the user according to the sleep posture state monitoring data and the physiological parameter values. However, this technology only relies on pressure sensors, environment and basic physiological parameters, but sleep quality is affected by more factors, such as brain wave activity, galvanic skin response, etc., which are not covered, resulting in poor sleep quality monitoring effect. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a sleep quality analysis system based on deep learning, which can more accurately evaluate sleep quality.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A sleep quality analysis system based on deep learning, comprising:

[0007] The multi-modal signal acquisition module is configured to acquire physiological signals and environmental signals of a target user; the physiological signals include electroencephalogram signals, electromyography signals and skin resistance signals; and the environmental signals include temperature and humidity signals and air quality signals.

[0008] The denoising module is configured to perform denoising processing on the physiological signals and the environmental signals to obtain denoised signals.

[0009] The feature extraction module is configured to perform feature extraction on the denoised signals to obtain signal features.

[0010] The feature processing module is configured to perform data cleaning, data filling and normalization processing on the signal features to obtain normalized features.

[0011] The evaluation module is configured to input the normalized features into a trained sleep quality evaluation model to obtain an evaluation result; the sleep quality evaluation model is constructed by using a deep learning network.

[0012] Preferably, the denoising module comprises:

[0013] The signal decomposition submodule is configured to respectively decompose the physiological signals and the environmental signals into sub-signals of different frequencies using basis functions.

[0014] The coefficient acquisition submodule is configured to acquire transform coefficients corresponding to the sub-signals of different frequencies.

[0015] The variance calculation submodule is configured to calculate the variance of the transform coefficients at each decomposition scale.

[0016] The model construction submodule is configured to construct a signal denoising model based on the variance of the transform coefficients.

[0017] The preprocessing submodule is configured to perform preprocessing on the physiological signals and the environmental signals using the signal denoising model to obtain the denoised signals.

[0018] Preferably, the model construction submodule comprises:

[0019] The standard deviation estimation unit is configured to estimate the standard deviation of abnormal noise according to the variance of the transform coefficients.

[0020] The threshold construction unit is configured to construct a coefficient removal threshold according to the standard deviation of the noise; wherein the coefficient removal threshold is: wherein t represents the coefficient removal threshold, σ0 represents the variance of the transform coefficients, σ p represents the standard deviation of abnormal noise, median(d p ) represents the median of the transform coefficients at the pth decomposition scale.

[0021] The denoising model construction unit is configured to construct a signal denoising model using the coefficients.

[0022] Preferably, the signal denoising model is:

[0023]

[0024] wherein, α represents a transform coefficient, w j,k represents a transform coefficient at the jth decomposition scale, m represents a convergence coefficient, sgn represents a sign function, w j,k represents an outlier-removed transform coefficient.

[0025] Preferably, the preprocessing sub-module comprises:

[0026] The coefficient setting unit is configured to set initial transform coefficients and convergence coefficients for the signal denoising model.

[0027] The processing unit is configured to process the transform coefficients using the signal denoising model with the setting completed to obtain the outlier-removed transform coefficients.

[0028] The reconstruction unit is configured to reconstruct the outlier-removed transform coefficients to obtain a reconstructed signal.

[0029] The signal-to-noise ratio calculation unit is configured to calculate a signal-to-noise ratio between the reconstructed signal and the original physiological signal and the original environmental signal.

[0030] The iteration unit is configured to reset the transform coefficients and the convergence coefficients when the signal-to-noise ratio is not in a preset range until the signal-to-noise ratio is maintained in the preset range.

[0031] Preferably, the signal features comprise physiological signal sub-features and environmental signal sub-features; the physiological signal sub-features comprise frequency band energy, power spectral density, entropy and complexity of an electroencephalogram signal, average amplitude, frequency and zero-crossing rate of an electromyography signal, conductivity change and skin galvanic response frequency of a skin resistance signal; and the environmental signal sub-features comprise average temperature and humidity and change rate of a temperature and humidity signal, carbon dioxide concentration and volatile organic compound level of an air quality signal.

[0032] Preferably, the feature processing module comprises:

[0033] The first traversal sub-module is configured to traverse all the signal features, detect repeated invalid data, and delete the detected repeated invalid data.

[0034] The second traversal sub-module is configured to traverse all the signal features, detect features with missing values in the data, wherein features with missing values exceeding 80% are deleted, and features with missing values less than 80% are filled based on a KNN algorithm.

[0035] An imbalance sub-module is configured to perform data imbalance processing on the feature data after data deletion and data filling based on the SMOTE algorithm.

[0036] A normalization sub-module is configured to perform normalization processing on the data after data imbalance processing based on a Z-score standardization method to obtain normalized features.

[0037] Preferably, the sleep quality evaluation model comprises a multi-modal input layer, a single-signal convolution layer, a shared feature fusion layer, a time series analysis layer, a feature attention mechanism layer, and an output evaluation layer; the multi-modal input layer is provided with an input channel corresponding to each signal type; the multi-modal input layer is connected with the single-signal convolution layer; the shared feature fusion layer is connected with the single-signal convolution layer; the shared feature fusion layer is configured to aggregate features from different single-signal convolution layers, integrate the features through a fusion algorithm, and obtain fused features; the time series analysis layer is connected with the shared feature fusion layer, and the time series analysis layer is configured to use a long short-term memory network to perform time series analysis on the fused features; the feature attention mechanism layer is connected with the shared feature fusion layer, and the feature attention mechanism layer is deployed with an attention mechanism to dynamically adjust the feature weights of different signals; the output evaluation layer is connected with the time series analysis layer, and the output evaluation layer is configured to comprehensively calculate the evaluation results according to the results of the time series analysis; the evaluation results comprise a sleep quality score.

[0038] Preferably, the calculation formula of the sleep quality score is as follows:

[0039]

[0040] wherein SQ is the sleep quality score, N is the number of signal features, a i is the attention weight of the i-th signal feature, f i is the value of the i-th signal feature, LSTM output is the output result of the LSTM layer, M is the number of environmental signal features, w j is the contribution coefficient of the j-th environmental feature, p j is the value of the j-th environmental signal feature, such as average temperature, humidity change, and air quality index, a is a control parameter for controlling the influence of single-signal feature aggregation on the total score, β is a control parameter for controlling the influence of time series dynamic features, γ is a control parameter for adjusting the proportion of environmental features in the total score, and δ is a bias term for adjusting the baseline of the final score.

[0041] According to the specific embodiments of the present application, the following technical effects are achieved:

[0042] The application provides a sleep quality analysis system based on deep learning, comprising: a multi-modal signal acquisition module, used for acquiring physiological signals and environmental signals of a target user; the physiological signals comprise brain wave signals, electromyography signals and skin resistance signals; the environmental signals comprise temperature and humidity signals and air quality signals; a denoising module, used for carrying out denoising processing on the physiological signals and the environmental signals to obtain denoised signals; a feature extraction module, used for carrying out feature extraction on the denoised signals to obtain signal features; a feature processing module, used for carrying out data cleaning, data filling and normalization processing on the signal features to obtain normalized features; an evaluation module, used for inputting the normalized features into a trained sleep quality evaluation model to obtain an evaluation result; and the sleep quality evaluation model is constructed through a deep learning network.The physiological signals (such as brain waves, electromyography and skin resistance) and the environmental signals (such as temperature and humidity and air quality) are simultaneously acquired, and multiple data sources provide a more comprehensive understanding of sleep quality.Different signals can complement each other, the physiological signals reflect the body condition, and the environmental signals can optimize the conditions, so that the sleep quality can be more accurately evaluated. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0044] Figure 1 The system structure schematic diagram provided for the embodiments of the present application is shown in DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0046] The purpose of the present application is to provide a sleep quality analysis system based on deep learning, which can accurately evaluate sleep quality.

[0047] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0048] Figure 1 The system structure schematic diagram provided for the embodiments of the present application is shown in Figure 1As shown, the present application provides a sleep quality analysis system based on deep learning, comprising:

[0049] A multi-modal signal acquisition module is configured to acquire physiological signals and environmental signals of a target user; the physiological signals include electroencephalogram signals, electromyography signals, and galvanic skin response signals; the environmental signals include temperature and humidity signals and air quality signals;

[0050] A denoising module is configured to perform denoising processing on the physiological signals and the environmental signals to obtain denoised signals;

[0051] A feature extraction module is configured to perform feature extraction on the denoised signals to obtain signal features;

[0052] A feature processing module is configured to perform data cleaning, data filling, and normalization processing on the signal features to obtain normalized features;

[0053] An evaluation module is configured to input the normalized features into a trained sleep quality evaluation model to obtain an evaluation result; the sleep quality evaluation model is constructed by a deep learning network.

[0054] Specifically, the physiological signal acquisition device of the present embodiment comprises:

[0055] An electroencephalogram (EEG) sensor, a high-precision EEG headset is selected, usually including multiple electrodes, is used to acquire electrical signals of brain activity.

[0056] An electromyography (EMG) sensor, a patch-type or wearable EMG sensor is used, which is attached to the surface of the muscle to detect electrical signals of muscle activity.

[0057] A galvanic skin response (GSR) sensor, a GSR sensor worn on the finger or wrist is selected, which is used to measure the change of skin conductivity and reflect the activity of the autonomic nervous system.

[0058] Further, the environmental signal acquisition device of the present embodiment comprises:

[0059] A temperature and humidity sensor, an integrated temperature and humidity sensor module is used, which can monitor the environmental temperature and humidity in real time.

[0060] An air quality sensor, a sensor module capable of detecting carbon dioxide concentration and volatile organic compounds (VOC) is selected to evaluate air cleanliness.

[0061] Further, the embodiment correctly installs physiological signal sensors at designated positions on the user's body, ensuring the accuracy of signal acquisition. Environmental sensors are placed in the user's sleep environment. Through the built-in acquisition module of the sensors, physiological and environmental signals are obtained in real time. Typically, these devices collect data at a certain sampling rate (such as hundreds of times per second). Using Bluetooth, Wi-Fi, or other wireless communication technologies, the collected signal data is transmitted to a central processing unit (such as a computer or mobile device). Demonstratively, the embodiment ensures that the timestamps of all signal data are synchronized for subsequent multi-modal data fusion and analysis.

[0062] Further, the embodiment stores the data transmitted to the central processing unit in time series format. After data storage, preliminary denoising is performed first, using filters such as low-pass filters and high-pass filters to remove noise and interference in the signal. Then, key features are extracted from the denoised signal, such as EEG band energy, EMG average amplitude, GSR conductivity changes, etc. Subsequently, missing data and outliers are processed to ensure data integrity and consistency, and feature data is normalized to facilitate subsequent model input and analysis. Finally, the embodiment inputs the preprocessed data into the trained deep learning model for sleep quality assessment. Based on the analysis results, real-time feedback and suggestions are provided to help users improve their sleep environment and habits.

[0063] Through the above steps, the system can effectively collect and process the user's physiological and environmental signals to provide a reliable data basis for sleep quality assessment.

[0064] Preferably, the denoising module comprises:

[0065] A signal decomposition submodule for decomposing the physiological signal and the environmental signal into sub-signals of different frequencies using basis functions, respectively;

[0066] A coefficient acquisition submodule for acquiring transform coefficients corresponding to the sub-signals of different frequencies;

[0067] A variance calculation submodule for calculating the variance of the transform coefficients at each decomposition scale;

[0068] A model construction submodule for constructing a signal denoising model based on the variance of the transform coefficients;

[0069] A preprocessing submodule for preprocessing the physiological signal and the environmental signal using the signal denoising model to obtain the denoised signal.

[0070] Preferably, the model construction submodule comprises:

[0071] a standard deviation estimation unit configured to estimate a standard deviation of the abnormal noise according to a variance of the transform coefficients;

[0072] a threshold construction unit configured to construct a coefficient removal threshold according to the standard deviation of the noise; wherein the coefficient removal threshold is: wherein t represents the coefficient removal threshold, σ0 represents the variance of the transform coefficients, σ p represents the standard deviation of the abnormal noise, median(d p ) represents a median value of the transform coefficients at the pth decomposition scale;

[0073] a denoising model construction unit configured to construct a signal denoising model by using the coefficient removal threshold.

[0074] Preferably, the signal denoising model is:

[0075]

[0076] wherein α represents the transform coefficient, w j,k represents the transform coefficient at the jth decomposition scale, m represents the convergence coefficient, sgn represents a sign function, and w ,k represents the transform coefficient after removing the abnormal value.

[0077] The present application is based on the principle of wavelet transform to transform the original signal into sub-signals of different frequencies. The signal denoising model is continuous at t, the function value is (1-α)t, and the first derivative is 1-α at t. The convergence speed can be adjusted by m, and the greater m is, the faster the convergence is. The shrinkage degree of the wavelet coefficient within ±t can be adjusted by α, so that the function can be transformed between the conventional hard threshold function and the soft threshold function, and the characteristics of the sequence signal can be better adapted to remove the noise interference perfectly.

[0078] Preferably, the preprocessing sub-module comprises:

[0079] a coefficient setting unit configured to set an initial transform coefficient and a convergence coefficient for the signal denoising model;

[0080] a processing unit configured to process the transform coefficient by using the signal denoising model after the setting is completed to obtain the transform coefficient after removing the abnormal value;

[0081] a reconstruction unit configured to reconstruct the transform coefficient after removing the abnormal value to obtain a reconstructed signal;

[0082] a signal-to-noise ratio calculation unit configured to calculate a signal-to-noise ratio between the reconstructed signal and the original physiological signal and the original environmental signal;

[0083] The iteration unit is configured to reset the transform coefficient and the convergence coefficient when the signal-to-noise ratio is not in the preset range until the signal-to-noise ratio remains in the preset range.

[0084] Preferably, the signal features include physiological signal sub-features and environmental signal sub-features; the physiological signal sub-features include: band energy, power spectral density, entropy and complexity of electroencephalogram signals, average amplitude, frequency and zero-crossing rate of electromyography signals, conductivity changes and galvanic skin response frequency of skin resistance signals; the environmental signal sub-features include: average temperature and humidity and change rate of temperature and humidity signals, carbon dioxide concentration and volatile organic compound level of air quality signals.

[0085] Specifically, the physiological signal sub-features of the embodiment are as follows:

[0086] Electroencephalogram (EEG), band energy, extract energy of α, β, θ, δ bands, for identifying sleep stages (such as REM, deep sleep).

[0087] Power spectral density (PSD), analyze power distribution of different frequency components, help identify sleep depth and quality.

[0088] Entropy and complexity, such as sample entropy, approximate entropy, for evaluating complexity and variability of electroencephalogram.

[0089] Electromyography (EMG):

[0090] Average amplitude and frequency, for detecting muscle activity level, helping to identify movement events (such as restless legs) in sleep.

[0091] Zero-crossing rate, for analyzing frequency characteristics of muscle activity.

[0092] Galvanic skin resistance (GSR):

[0093] Conductivity changes, reflecting autonomic nervous system activity, helping to identify stress response and wakefulness.

[0094] Galvanic skin response frequency, for assessing emotional and stress levels.

[0095] Further, the environmental signal sub-features of the embodiment are as follows:

[0096] Temperature and humidity signals:

[0097] Average temperature and humidity, for evaluating comfort of sleep environment.

[0098] Change rate, analyze fluctuations in temperature and humidity, identify environmental changes that may affect sleep.

[0099] Air quality signals:

[0100] Carbon dioxide concentration, high concentration can cause sleep quality to decline.

[0101] Volatile organic compounds (VOC) levels, used to assess air cleanliness.

[0102] Further, the embodiment identifies different sleep stages (such as light sleep, deep sleep, REM) by analyzing EEG band energy and PSD, thereby evaluating sleep structure. Sleep continuity is evaluated by identifying movement events and wake times in sleep using EMG features. Stress response in sleep is evaluated by analyzing autonomic nervous system activity through GSR features. The influence of sleep environment comfort on sleep quality is evaluated by combining temperature and humidity and air quality features.

[0103] The above features of the embodiment are analyzed by a deep learning model, which can provide a comprehensive evaluation of sleep quality, help identify key factors affecting sleep, and provide personalized improvement suggestions.

[0104] Preferably, the feature processing module comprises:

[0105] The first traversal submodule is configured to traverse all signal features, detect repeated and invalid data, and delete the detected repeated and invalid data;

[0106] The second traversal submodule is configured to traverse all signal features, detect features with missing values in the data, and delete features with missing values exceeding 80%, and fill in features with missing values less than 80% based on the KNN algorithm.

[0107] The imbalance submodule is configured to perform data imbalance processing on the feature data after data deletion and data filling based on the SMOTE algorithm.

[0108] The normalization submodule is configured to perform normalization processing on the data imbalance processed data based on the Z-score standardization method to obtain the normalized features.

[0109] Preferably, the sleep quality evaluation model comprises a multi-modal input layer, a single-signal convolution layer, a shared feature fusion layer, a time series analysis layer, a feature attention mechanism layer, and an output evaluation layer; the multi-modal input layer is provided with an input channel corresponding to each signal type; the multi-modal input layer is connected with the single-signal convolution layer; the shared feature fusion layer is connected with the single-signal convolution layer; the shared feature fusion layer is used to aggregate features from different single-signal convolution layers, integrate the features through a fusion algorithm, and obtain fused features; the time series analysis layer is connected with the shared feature fusion layer, and the time series analysis layer is used to perform time series analysis on the fused features using a long short-term memory network; the feature attention mechanism layer is connected with the shared feature fusion layer, and the feature attention mechanism layer is deployed with an attention mechanism to dynamically adjust the feature weights of different signals; the output evaluation layer is connected with the time series analysis layer, and the output evaluation layer is used to comprehensively calculate the evaluation results according to the results of the time series analysis; and the evaluation results comprise a sleep quality score.

[0110] Specifically, the model structure of the embodiment comprises:

[0111] The multi-modal input layer is designed with separate input channels for each signal type (EEG, EMG, GSR, temperature and humidity, air quality), so that the features of each signal can independently enter the model.

[0112] The single-signal convolution layer uses a convolutional neural network (CNN) for each input channel to extract local features that can capture data patterns of the signal, such as time features, frequency domain changes, etc.

[0113] The shared feature fusion layer aggregates features from different signal convolution layers and integrates the features through a fully connected network (FCN) or a fusion algorithm (such as an attention mechanism) to capture the mutual relationship between signals.

[0114] The time series analysis layer uses a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to perform time series analysis on the fused features, which can capture the time dynamic changes and trends of the signal.

[0115] The feature attention mechanism layer deploys an attention mechanism to dynamically adjust the feature weights of different signals, highlighting and identifying the signal features that have the most impact on sleep quality.

[0116] The output evaluation layer performs comprehensive calculation of the evaluation results through a multi-layer perceptron (MLP) to output a sleep quality score or classification (such as good, average, poor).

[0117] Further, each signal feature of the embodiments enters through its corresponding input layer. The design of different input channels allows the model to optimize processing for different data structures, such as one-dimensional time series and two-dimensional spectrograms. Each single signal convolutional layer extracts local features of the respective signal. This results in each convolutional layer outputting a set of feature maps reflecting the local characteristics of the signals. The features output by multiple convolutional layers are integrated in a shared feature fusion layer. This fusion helps to reveal potential combined effects between signals, such as the combination of EEG and EMG can reveal the relationship between sleep and movement. The fused feature sequence is input to the LSTM layer, which is suitable for mining information that changes over time, and is important for understanding how actions, wake-up cycles, and environmental changes affect sleep. By calculating the attention weights of different features (dynamic weight distribution), it is identified that some signals may be more important to the final sleep quality assessment in a specific context, improving the accuracy of the assessment results. The final comprehensive features are used by the MLP to obtain the sleep quality score, which provides a quantitative assessment of the user's overall sleep status.

[0118] Further, in the model training process, to achieve personalized improvement solutions, a large amount of historical sleep data of individual users is first collected, including physiological signals, environmental signals, and users' daily routines and living habits. Through deep learning models, these multi-modal data are processed and analyzed to capture the sleep performance characteristics of individual users under different environmental conditions and different routines. The model uses a task-specific loss function to optimize the accuracy of personalized recommendations. At the same time, by implementing the continuous learning ability of the model, the user's personal profile can be updated after new data input by the user, identifying sleep problems under specific conditions. Then, combined with classification analysis and regression analysis, targeted recommendations are generated to truly match the individual needs of the user. For example, when the user is often affected by environmental noise or air quality changes, the system may recommend installing an air purification device or using a white noise machine. At the same time, the system can also recommend adjusting routines based on long-term sleep habit data, such as gradually changing bedtime, to better adapt to the user's physiological rhythm. This makes the improvement solution not only scientific and individually adaptive, but also dynamically adjustable and continuously optimized.

[0119] Preferably, the calculation formula of the sleep quality score is:

[0120]

[0121] where SQ is the sleep quality score, N is the number of signal features, a i is the attention weight of the i-th signal feature, f i is the value of the i-th signal feature, LSTM output is the output result of the LSTM layer, M is the number of environmental signal features, w jContribution coefficient of the jth environmental feature, p j Value of the jth environmental signal feature, such as average temperature, humidity change, and air quality index, α controls the influence of the aggregation of single signal features on the total score, β controls the influence of time series dynamic features, γ adjusts the proportion of environmental features in the total score, and δ is a bias term used to adjust the baseline of the final score.

[0122] Specifically, the parameters of the above formula in this embodiment are explained as follows:

[0123] SQ (Sleep Quality): Sleep quality score, final output result, usually a score of 0 to 100, representing the overall sleep quality, from poor to excellent.

[0124] N: Number of signal features, usually the total number of different features extracted from multi-modal signals.

[0125] a i : Attention weight of the ith signal feature, dynamically adjusted by the feature attention mechanism layer according to the contribution of different features to sleep quality.

[0126] f i : Value of the ith signal feature, extracted from the shared feature fusion layer, which may include, for example, EEG frequency band energy, EMG amplitude, GSR conductivity change, etc.

[0127] LSTM output : Output result of the LSTM layer, which captures the time dynamic change information of the fusion features, reflecting the influence of time characteristics on sleep quality.

[0128] M: Number of environmental signal features, such as temperature and humidity and air quality related features.

[0129] w j : Contribution coefficient of the jth environmental feature, evaluating the influence of the environmental feature on overall sleep quality, usually based on prior knowledge or data-driven adjustment.

[0130] p j : Value of the jth environmental signal feature, such as average temperature, humidity change, and air quality index.

[0131] α, β, γ and δ are all model hyperparameters, used to control the influence of each part on the final sleep quality score. The specific values can be adjusted through model training and validation set, specifically:

[0132] α: Controls the influence of the aggregation of single signal features on the total score.

[0133] β: Controls the influence of time series dynamic features.

[0134] gamma: adjust the proportion of environmental characteristics in the total score.

[0135] delta: bias term, used to adjust the baseline of the final score.

[0136] Optionally, in addition to the sleep quality score, the evaluation results also include sleep stage analysis, number of awakenings, sleep efficiency, sleep latency, deep sleep proportion, environmental impact factor analysis, and personalized improvement suggestions. Sleep stage analysis provides the time distribution of the user in different sleep stages (such as light sleep, deep sleep, REM); the number of awakenings and sleep efficiency reflect the continuity and quality of sleep; sleep latency refers to the time required to fall asleep; deep sleep proportion is a key indicator of restorative sleep. Environmental impact factor analysis reveals the impact of temperature, humidity, and air quality on sleep. Personalized improvement suggestions are based on the evaluation results to provide specific measures for users to optimize their sleep environment and habits. Through these comprehensive information, users can have a comprehensive understanding of their own sleep condition and take effective measures to improve it.

[0137] The beneficial effects of the present application are as follows:

[0138] (1) The present application simultaneously collects physiological signals (such as brain waves, electromyography, skin resistance) and environmental signals (such as temperature, humidity, air quality), providing a more comprehensive understanding of sleep quality from multiple data sources. Different signals can complement each other, with physiological signals reflecting physical condition and environmental signals optimizing conditions, thus more accurately assessing sleep quality.

[0139] (1) The denoising module of the present application effectively reduces false signals introduced by the environment and equipment, ensuring the accuracy of the data. The feature extraction module can extract key features from complex signals, actively capturing information in time and frequency domains, and improving analysis results.

[0140] (3) The deep learning network of the present application can adaptively learn feature relationships and capture potential patterns according to the complexity of multi-modal input data, which is difficult to achieve in traditional models. Deep learning has strong non-linear mapping ability and can handle complex data correlations, improving the flexibility and accuracy of the model.

[0141] (4) The present application combines deep learning technology, and the system can dynamically process data and provide real-time sleep quality scores and suggestions, thereby guiding users to improve their sleep. During model training, personalized improvement plans can be generated for specific user data, such as adjusting sleep environment and optimizing routines.

[0142] (6) The model of the present application is trained on a large amount of historical data, reducing the dependence on human experience for results and improving the reliability and accuracy of the results. Through continuous collection of user feedback and effect monitoring, the system can continuously iterate and optimize.

[0143] The various embodiments described in this specification are presented for the purpose of illustrating the principles of the present application and its best mode of operation. Each of the embodiments described in this specification has been provided for the purpose of illustration and is not intended to limit the application.

[0144] The principles and operation of the present application have been explained so far with the help of specific examples. The above examples are only for the purpose of helping to understand the method of the present application and its core idea. At the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation and application range. In view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1.A deep learning-based sleep quality analysis system, characterized by, The method comprises the following steps: A multi-modal signal acquisition module is used to acquire physiological signals and environmental signals of a target user; the physiological signals include brain wave signals, electromyography signals, and skin resistance signals; the environmental signals include temperature and humidity signals and air quality signals; A denoising module is used to perform denoising processing on the physiological signals and the environmental signals to obtain denoised signals; A feature extraction module is used to perform feature extraction on the denoised signals to obtain signal features; A feature processing module is used to perform data cleaning, data filling, and normalization processing on the signal features to obtain normalized features; An evaluation module is used to input the normalized features into a trained sleep quality evaluation model to obtain an evaluation result; the sleep quality evaluation model is constructed by using a deep learning network; The sleep quality evaluation model comprises a multi-modal input layer, a single-signal convolution layer, a shared feature fusion layer, a time series analysis layer, a feature attention mechanism layer, and an output evaluation layer; the multi-modal input layer is provided with an input channel corresponding to each type of signal; the multi-modal input layer is connected with the single-signal convolution layer; the shared feature fusion layer is connected with the single-signal convolution layer; the shared feature fusion layer is used to aggregate features from different single-signal convolution layers, integrate the features by using a fusion algorithm, and obtain fused features; The time series analysis layer is connected with the shared feature fusion layer, and the time series analysis layer is used to perform time series analysis on the fused features by using a long short-term memory network; the feature attention mechanism layer is connected with the shared feature fusion layer, and the feature attention mechanism layer is deployed with an attention mechanism to dynamically adjust the feature weights of different signals; the output evaluation layer is connected with the time series analysis layer, and the output evaluation layer is used to comprehensively calculate the evaluation result according to the result of the time series analysis; the evaluation result includes a sleep quality score; The calculation formula of the sleep quality score is: ； wherein, is the sleep quality score, is the number of signal features, is the attention weight of the th signal feature, is the value of the th signal feature, is the output result of the LSTM layer, is the number of environmental signal features, is the contribution coefficient of the th environmental feature, is the value of the th environmental signal feature, such as average temperature, humidity change, and air quality index, a is the control of the influence of single signal feature aggregation on the total score, β is the influence of the time series dynamic feature, γ is the adjustment of the proportion of environmental features in the total score, and δ is the bias term for adjusting the baseline of the final score. 2.The deep learning-based sleep quality analysis system of claim 1, wherein, The denoising module comprises: A signal decomposition sub-module is used to respectively decompose the physiological signals and the environmental signals into sub-signals of different frequencies by using a base function; A coefficient acquisition sub-module is used to acquire transform coefficients corresponding to the sub-signals of different frequencies; A variance calculation sub-module is used to calculate the variance of the transform coefficients at each decomposition scale; A model construction sub-module is used to construct a signal denoising model based on the variance of the transform coefficients; A preprocessing sub-module is used to pre-process the physiological signals and the environmental signals by using the signal denoising model to obtain the denoised signals. 3.The deep learning-based sleep quality analysis system of claim 2, wherein, The model construction sub-module comprises: A standard deviation estimation unit is used to estimate the standard deviation of abnormal noise according to the variance of the transform coefficients; a threshold construction unit configured to construct a coefficient removal threshold according to a standard deviation of the noise; wherein the coefficient removal threshold is: ; wherein, denotes the coefficient removal threshold, denotes a variance of the transform coefficients, denotes a standard deviation of the abnormal noise, denotes a median value of the transform coefficients at the pth decomposition scale; A denoising model construction unit is used to construct a signal denoising model by using the coefficient removal threshold. 4.The deep learning-based sleep quality analysis system of claim 3, wherein, The signal denoising model is: ； wherein, denotes a transform coefficient, denotes a transform coefficient at the j-th decomposition scale, denotes a convergence coefficient, denotes a sign function, denotes a transform coefficient with outliers removed. 5.The deep learning-based sleep quality analysis system of claim 4, wherein, The preprocessing sub-module comprises: A coefficient setting unit is used to set initial transform coefficients and convergence coefficients for the signal denoising model; A processing unit is used to process the transform coefficients by using the signal denoising model with the setting completed to obtain transform coefficients with abnormal values removed; A reconstruction unit is used to reconstruct the transform coefficients with abnormal values removed to obtain reconstructed signals; A signal-to-noise ratio calculation unit is configured to calculate a signal-to-noise ratio between the reconstructed signal and the original physiological signal and the environmental signal. An iteration unit is configured to reset the transform coefficient and the convergence coefficient when the signal-to-noise ratio is not in the preset range until the signal-to-noise ratio remains in the preset range. 6.The deep learning-based sleep quality analysis system of claim 1, wherein, The signal features include physiological signal sub-features and environmental signal sub-features; the physiological signal sub-features include band energy, power spectral density, entropy and complexity of brain wave signals, average amplitude, frequency and zero-crossing rate of electromyography signals, conductivity change of skin resistance signals and skin electric response frequency; and the environmental signal sub-features include average temperature and humidity and change rate of temperature and humidity signals, carbon dioxide concentration and volatile organic compound level of air quality signals. 7.The deep learning-based sleep quality analysis system of claim 1, wherein, The feature processing module includes: A first traversal sub-module is configured to traverse all signal features, detect repeated invalid data, and delete the detected repeated invalid data; A second traversal sub-module is configured to traverse all signal features, detect features with missing data in the data, delete features with missing values exceeding 80%, and fill in features with missing values less than 80% based on a KNN algorithm; An imbalance sub-module is configured to perform data imbalance processing on the feature data after data deletion and data filling based on an SMOTE algorithm; A normalization sub-module is configured to perform normalization processing on the data after data imbalance processing based on a Z-score standardization method to obtain the normalized features.

Images